NASA Earth Observation Knowledge Graph

• NASA EO‑KG is a public heterogeneous knowledge graph that connects Earth observation datasets, publications, science keywords, instruments, and projects.
• It supports dual methodologies by serving as an offline source for publication–dataset citation in retrieval benchmarks and as a multi‐relational graph for learning novel dataset pairings.
• Key advancements include robust dataset-publication linkages, significant retrieval performance gains, and a framework for integrating diverse metadata in Earth observation research.

The NASA Earth Observation Knowledge Graph (NASA EO‑KG) is a public, heterogeneous knowledge graph for Earth observation that links datasets, publications, science keywords, instruments, platforms, projects, and data centers. In recent work it has been used in two distinct but complementary ways: as an authoritative offline source of publication–dataset usage for search and retrieval, and as a multi-relational graph for learning novel dataset pairings and grounding hypothesis generation (Yu et al., 2 Jul 2026, Ghazanfari et al., 2 Jul 2026). In the search setting, its operative core is the publication–dataset citation subgraph, especially USES_DATASET edges from peer-reviewed publications to NASA Common Metadata Repository (CMR) datasets; in the hypothesis-generation setting, a broader typed graph is used directly for heterogeneous graph learning (Yu et al., 2 Jul 2026, Ghazanfari et al., 2 Jul 2026).

1. Definition and scope

NASA EO‑KG is described as the NASA Earth Observation Knowledge Graph (NASA EO‑KG), released as a dataset on Hugging Face, and in later work as a public GraphML resource with 150,351 nodes and 436,203 edges spanning seven node types and nine typed relations (Yu et al., 2 Jul 2026, Ghazanfari et al., 2 Jul 2026). Conceptually, it links peer-reviewed publications to NASA Earth observation datasets and also organizes dataset metadata through science keywords, platforms, instruments, projects, and data centers (Ghazanfari et al., 2 Jul 2026).

In the agentic-search formulation, NASA EO‑KG is treated as an existing, external resource rather than as an online reasoning substrate. The paper is explicit that the runtime system does not query the graph; instead, the graph is used offline to extract high-provenance USES_DATASET edges, which become “silver labels” for supervised retrieval (Yu et al., 2 Jul 2026). This makes a crucial distinction: the graph is not synonymous with the deployed search agent, even though it underwrites that agent’s effectiveness.

A common misunderstanding is therefore that NASA EO‑KG is only a publication–dataset graph. That statement is accurate for the way the search paper uses it, but incomplete as a description of the resource itself. In the hypothesis-generation setting, EO‑KG is a broader heterogeneous graph whose nodes include Publication, Dataset, ScienceKeyword, Instrument, Platform, Project, and DataCenter (Ghazanfari et al., 2 Jul 2026).

2. Graph structure and representation

The broader EO‑KG schema reported for hypothesis generation includes the following node inventory and counts (Ghazanfari et al., 2 Jul 2026):

Node type	Count	Attributes noted
Publication	138,704	doi, title, year, authors, abstract
Dataset	8,058	shortName, longName, DAAC, CMR ID, temporal extent
ScienceKeyword	1,609	keyword name, hierarchical subcategory
Instrument	921	not detailed beyond type
Platform	455	not detailed beyond type
Project	415	not detailed beyond type
DataCenter	189	not detailed beyond type

The nine typed relations reported in that same graph are CITES, HAS_APPLIEDRESEARCHAREA, USES_DATASET, HAS_SCIENCEKEYWORD, HAS_PLATFORM, HAS_DATASET, OF_PROJECT, HAS_INSTRUMENT, and HAS_SUBCATEGORY (Ghazanfari et al., 2 Jul 2026). In the narrower search usage, however, the relevant schema is essentially reduced to Publication, Dataset, and USES_DATASET(Publication, Dataset) (Yu et al., 2 Jul 2026).

Formally, the search paper relies on a bipartite citation subgraph. Let $\mathcal{P}$ be publications, $\mathcal{D}$ be NASA CMR datasets, and $$\mathcal{E}_{\text{uses}} \subset \mathcal{P} \times \mathcal{D}$$ the set of USES_DATASET edges. Each edge $(p,d) \in \mathcal{E}_{\text{uses}}$ means publication $p$ used dataset $d$ (Yu et al., 2 Jul 2026). Dataset nodes are already reconciled with CMR collection identifiers, and publication nodes are already linked to NASA GES DISC literature with citation metadata (Yu et al., 2 Jul 2026).

This split between the full heterogeneous graph and the publication–dataset citation slice is central to understanding the resource. In one usage mode, EO‑KG behaves like a typed, multi-relational graph suitable for link prediction. In the other, it acts as an authoritative provenance layer from which text-only retrieval benchmarks can be derived.

3. NASA‑EO‑Bench and supervised retrieval from USES_DATASET

NASA‑EO‑Bench is the principal benchmark derived from EO‑KG for data discovery (Yu et al., 2 Jul 2026). Its construction begins with the top 10,636 publications by citation count from NASA EO‑KG. For each publication $p$, the graph defines a ground-truth dataset set

$$\mathcal{G}_p = \{ d \in \mathcal{D} \mid (p,d) \in \mathcal{E}_{\text{uses}} \}.$$

Two task-based natural-language queries are then generated per publication using GPT‑5.4 conditioned only on the abstract; the model is not told which datasets were actually used (Yu et al., 2 Jul 2026). After filtering, this yields 21,272 task-based queries and 47,654 query–dataset positive pairs over a corpus of 8,058 NASA CMR datasets (Yu et al., 2 Jul 2026).

The split is by publication to avoid leakage: 17,038 queries / 38,078 pairs for training and 4,234 queries / 9,576 pairs for test, stratified by cited_by_count (Yu et al., 2 Jul 2026). The benchmark therefore preserves the literature-grounded linkage between a research need and the datasets actually used in a paper.

Evaluation is defined against the KG-derived ground-truth set $\mathcal{G}_q$. The benchmark uses Recall@K, MRR, and MAP, with

$$R@K(q) = \frac{|\mathcal{R}_q^{(K)} \cap \mathcal{G}_q|}{|\mathcal{G}_q|},$$

and macro-averaged recall and reciprocal-rank metrics over all queries (Yu et al., 2 Jul 2026). Because the labels arise from actual publication usage, they encode strong provenance; because they are citation-driven rather than exhaustively adjudicated, they remain silver rather than gold labels.

This use of EO‑KG is distinctive. The graph’s information is injected not as runtime graph features, but as supervision. The neural retrievers never traverse the graph at inference time; instead, they learn from large-scale query–dataset relevance pairs extracted from the graph.

4. Runtime search pipeline and the non-runtime role of the graph

The deployed search pipeline has three stages: Router, Hybrid retrieval, and Agentic reranking (Yu et al., 2 Jul 2026). The Router uses LLM routing to send queries to official NASA services such as Harmony, Science Discovery Engine, Worldview, and Giovanni when intent matches those tools. The Hybrid retrieval stage operates over the 8,058-dataset text corpus derived from CMR, representing each dataset by shortName + longName + abstract, plus additional metadata for BM25 (Yu et al., 2 Jul 2026). The Agentic reranking stage then sees the query and top‑$\mathcal{D}$0 candidate dataset descriptions, and may call general web search and arXiv search; in the deployed system, SDE search is also available but was not used in the reported experiments (Yu et al., 2 Jul 2026).

The neural retrieval stack includes BM25, cosine similarity over a NASA domain-adapted sentence transformer, a fine-tuned cosine bi-encoder, and NN‑SSC, a neural score-correction MLP. NN‑SSC takes L2-normalized query and dataset embeddings, concatenates them into $\mathcal{D}$1, and applies an MLP with architecture

$\mathcal{D}$2

with ReLU hidden layers and sigmoid output (Yu et al., 2 Jul 2026). Hybrid score fusion combines a normalized lexical score $\mathcal{D}$3 and neural score $\mathcal{D}$4 as

$\mathcal{D}$5

with $\mathcal{D}$6 set analytically from standalone training performance (Yu et al., 2 Jul 2026).

The reported gains are large. The Cosine (base) baseline achieves R@10 = 0.0755 and MRR = 0.0538, whereas NN‑SSC + BM25 (Hybrid) achieves R@10 = 0.4275 and MRR = 0.2918, a more than 5× improvement in both metrics (Yu et al., 2 Jul 2026). On a stratified $\mathcal{D}$7 subset, agentic reranking further improves the hybrid baseline from MRR ≈ 0.302 to 0.388, with MAP rising from 0.260 to 0.323; the paper describes this as approximately 28% relative improvement over the baseline MRR (Yu et al., 2 Jul 2026).

The decisive point is methodological rather than architectural: NASA EO‑KG is used once, offline, at benchmark-construction time. BM25, dense retrieval, and the agentic reranker all operate over text and external APIs rather than graph queries (Yu et al., 2 Jul 2026). This directly rebuts the assumption that the search system is a runtime KG-reasoning engine.

5. Heterogeneous graph learning and hypothesis generation

A separate line of work uses the broader EO‑KG directly as a heterogeneous graph for scientific discovery (Ghazanfari et al., 2 Jul 2026). Starting from the public GraphML resource, the authors derive dataset–dataset co-usage pairs from Publication → Dataset edges: for each publication with dataset set $\mathcal{D}$8, every unordered pair in $\mathcal{D}$9 becomes a co-usage observation tagged with publication year (Ghazanfari et al., 2 Jul 2026). The resulting co-usage graph contains 2,088 dataset nodes with at least one co-usage edge, 23,490 unique dataset–dataset co-usage pairs, and a total co-usage weight of 64,930; edge weights are heavy-tailed, with maximum weight 1,051, median degree 11, and maximum degree 314 (Ghazanfari et al., 2 Jul 2026).

For learning, publication nodes are removed to avoid message-passing leakage, and the training graph retains Dataset, ScienceKeyword, Instrument, Platform, and Project nodes with six relations: co_usage, has_platform, has_keyword, has_instrument, of_project, and has_subcategory (Ghazanfari et al., 2 Jul 2026). Dataset nodes are initialized with 768-dimensional SPECTER2 embeddings of dataset abstracts; non-dataset nodes receive learnable 768-dimensional embeddings, and all node types are projected into a shared latent dimension 128 (Ghazanfari et al., 2 Jul 2026).

The link predictor is a two-layer heterogeneous GraphSAGE. For relation $$\mathcal{E}_{\text{uses}} \subset \mathcal{P} \times \mathcal{D}$$0 and node $$\mathcal{E}_{\text{uses}} \subset \mathcal{P} \times \mathcal{D}$$1, the relation-specific message is

$$\mathcal{E}_{\text{uses}} \subset \mathcal{P} \times \mathcal{D}$$2

and the update is

$$\mathcal{E}_{\text{uses}} \subset \mathcal{P} \times \mathcal{D}$$3

with ReLU and dropout (Ghazanfari et al., 2 Jul 2026). Two scoring heads are used,

$$\mathcal{E}_{\text{uses}} \subset \mathcal{P} \times \mathcal{D}$$4

trained with binary cross-entropy and degree-biased negative sampling (Ghazanfari et al., 2 Jul 2026).

This graph-learning stage feeds a three-agent LLM pipeline. The GNN ranks predicted-novel dataset pairs among a 1,475-dataset evaluation universe, and the top 200 novel pairs become candidates for Agent 1 filtering (Ghazanfari et al., 2 Jul 2026). Agent 2 turns selected pairs into structured research hypotheses, and Agent 3 evaluates them. Applied to 1,475 NASA datasets, the system produces 160 hypotheses spanning ecohydrology, glaciology, aerosol–cloud interactions, vegetation phenology, and stratospheric chemistry (Ghazanfari et al., 2 Jul 2026). Agent 1 ratings show that GNN-predicted novel pairs are nearly as plausible as held-out real co-usages: under GPT‑5.2, real co-usages score 4.57 in plausibility and predicted-novel pairs 4.38; under Claude Sonnet 4.6, the corresponding scores are 4.63 and 4.34 (Ghazanfari et al., 2 Jul 2026).

Here, unlike the search system, EO‑KG is not merely a label source. It is the structural object over which graph learning is performed, and the dataset metadata carried by graph nodes is passed directly into the LLM agents as grounding context.

6. Limitations, biases, and broader architectural implications

Both major usages of NASA EO‑KG expose structural limitations. In NASA‑EO‑Bench, labels are citation-driven, so coverage bias and popular dataset bias are unavoidable: many relevant but uncited products never appear as positives, and frequently used datasets such as MODIS, Landsat, and GPM IMERG are over-represented (Yu et al., 2 Jul 2026). Citation also does not exhaust relevance: a paper may use one of several scientifically equivalent datasets, so non-cited alternatives are counted as false positives. The search paper further notes potential label leakage in agentic evaluation, because a model with web and arXiv access might recover the source paper and reconstruct its dataset list (Yu et al., 2 Jul 2026).

In the graph-learning setting, the co-usage structure is heavy-tailed, publication leakage must be controlled by removing publication nodes, and the GNN uses binary rather than weighted co-usage supervision, which limits direct modeling of how strongly a pairing is established in the literature (Ghazanfari et al., 2 Jul 2026). Temporal information is used for splitting but not by a temporal GNN (Ghazanfari et al., 2 Jul 2026).

These limitations motivate the most plausible future directions already suggested by the papers themselves. The search paper explicitly points toward expanding EO‑KG entity and relation types to include variables, platforms, instruments, processing levels, and services such as Worldview, Giovanni, and Harmony, and toward possible runtime KG integration through live graph queries over publications and USES_DATASET edges (Yu et al., 2 Jul 2026). Related geospatial KG work suggests a standards-based route for such enrichment: RDF/OWL graphs with GeoSPARQL for geometry, OWL‑Time for temporal structure, SOSA/SSN for observations, QUDT for quantities, and PROV‑O for lineage (Zhu, 2024). Large-scale geo-KG systems such as KnowWhereGraph further suggest that a discrete global grid, precomputed spatial relations, and modular ontologies can support integration across environmental, hazard, and socio-economic layers at very large scale (Zhu et al., 19 Feb 2025).

Within NASA specifically, earlier Science Mission Directorate KG work had already framed knowledge graphs as a basis for dataset search engines across SMD domains, including Earth Science (Timmer et al., 2023). The recent EO‑KG literature narrows that general ambition into two concrete operational patterns: offline label extraction for search, and heterogeneous graph learning for discovery. Taken together, these uses establish NASA EO‑KG not as a single application, but as a reusable semantic asset whose publication–dataset provenance, typed metadata structure, and graph topology can support retrieval, recommendation, and hypothesis generation at different points in the Earth observation workflow.